Automatic Recognition and Generation of SNOMED codes in Medical Documents

Details

Supervisors

Vanderdonckt, Jean

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enManual SNOMED CT coding of clinical documents burdens physi- cians with a time-consuming, error-prone task that existing tools handle poorly for non-English text. At the Cliniques universitaires Saint-Luc in Brussels, ophthalmologists face this challenge daily: their French-language consultation notes are inadequately supported by the Epic EHR’s built-in coding assistance, motivating the need for an automated solution. This thesis implements and rigorously compares four computational approaches to automatic SNOMED CT entity linking: a three-tier dic- tionary with fuzzy matching (M1), a SapBERT bi-encoder with FAISS retrieval (M2), an LLM-driven named entity recognition pipeline grounded by SNOMED vector search (M3), and a hybrid architecture combining LLM span detection with tiered dictionary lookup (M4). On the SNOMED CT Entity Linking Challenge 1.2.1 benchmark (272 MIMIC-IV discharge summaries, 75,491 gold annotations), M1 achieves an mIoU of 0.3601, competitive with the challenge winner (0.420), while LLM-based methods score near zero (0.016–0.040). This apparent under- performance is explained by what this thesis identifies as the annotation density paradox: the benchmark exhaustively labels every medical term (∼340 per note), while LLM methods selectively extract clinically mean- ingful entities (∼35 per note). When measured independently of extraction density, M4 achieves the highest MRR (0.68) and a Hits@5 of 0.73 across all methods. The French clinical evaluation reveals a complete reversal of rankings. On six anonymised ophthalmological PDFs from Saint-Luc, LLM-based methods outperform dictionary approaches: M3 and M4 identify 43.8% of physician-validated diagnoses versus 37.5% for M1. A targeted prompt improvement explicitly instructing the model to capture systemic comor- bidities raises M4 to 56.3% (M4*), surpassing both GPT-3.5 and Claude Sonnet 4.6 without retrieval grounding, a 12.5-point gain requiring no model retraining. This advantage is confirmed on 36 structured French clinical cases (270 gold annotations), where M4 achieves the highest top-5 hit rate (42%). These results demonstrate that benchmark performance alone is an insufficient basis for clinical deployment decisions. For Saint-Luc, M4* of- fers the best trade-off between coverage, concept precision, and SNOMED validity guarantees, all without transmitting patient data to external servers.